The Use of Artificial Neural Networks to Describe and Predict the Presence of Harmful Algae in the Indian River Lagoon, Florida

Harmful algal blooms are a natural phenomenon of growing global concern. Dense blooms of single celled phytoplankton can have wide reaching effects on both the aquatic ecosystem and surrounding economies. This study constructed artificial neural network models of the northern Indian River Lagoon, Florida, using an existing dataset. Models attempted to both describe and predict chlorophyll a, as an indicator of total algal biomass, or Pyrodinium bahamense, a dinoflagellate known to bloom and produce the paralytic shellfish toxin saxitoxin in the lagoon. Descriptive models used current data while predictive models used time-lagged data as input. Further analyses were conducted on the best fitting descriptive models of chlorophyll a and P. bahamense in an attempt to elucidate driving factors of phytoplankton density within the ecosystem. Water samples were collected bimonthly for five years from six fixed sites in the northern Indian River Lagoon; a variety of environmental and hydrological parameters were collected and chemical and biological analyses done for each sample. Additional descriptive and meteorological data were collected or calculated for each site and added to other input variables. The dataset analyzed contained 645 samples, with at least 11 parameters recorded for each. vii Models of total chlorophyll a were relatively successful in describing absolute values and trends, and the predictive model (NMSE = 0.135, r = 0.933) was slightly more accurate than the descriptive (NMSE = 0.167, r = 0.913). Further analysis using metadata from the best descriptive model, known as “gray box” analyses, indicated that total phosphorus had a relatively large impact on overall chlorophyll a content in the water column. Models of P. bahamense attempted to describe or predict varying descriptors of density, including absolute density, density in known positive samples, relative density (high, medium, low) in known positive samples, and presence/absence. Only presence/absence classification models were relatively successful in describing or predicting P. bahamense density; descriptive models were accurate for 78.9% of samples while predictive models were accurate for 73% of samples. Further analysis of metadata from the best descriptive model offered very little insight beyond factors known to affect phytoplankton growth in laboratory based enrichment experiments